Temporal Query Alignment for ICU Deterioration Forecasting with Irregular Signals and Delayed Clinical Documentation

Abstract

Physiologic measurements are sampled irregularly, laboratory values arrive after processing delays, interventions alter both present and future patient state, and clinical notes summarize observations through language whose timing may not match the moment of bedside assessment. These properties make deterioration modeling a problem of temporal reasoning under partial observability rather than a straightforward exercise in classification over tabular data. This paper develops a framework that treats ICU forecasting as temporal query alignment over heterogeneous evidence. At each decision time, the model issues a structured query about future adverse events and retrieves support from irregular numeric trajectories, intervention context, and delayed textual documentation. The framework emphasizes state construction under uncertain document timing, retrieval of clinically relevant latent factors, and score formation under severe class imbalance. It introduces a query-conditioned state-space model, a document-alignment mechanism that separates filing time from latent observation time, and an optimization strategy that couples ranking performance with reliability-sensitive calibration. The analysis also addresses intervention entanglement, event-horizon ambiguity, missingness as a generated process, and deployment constraints such as alert budgets, support drift, and streaming inference. The broader argument is that useful ICU deterioration systems are best understood as temporal information-retrieval engines over evolving patient records, where success depends on how well the model aligns evidence across modalities and times before assigning probability to near-future decline